Guided Rotational Graph Embeddings for Error Detection in Noisy Knowledge Graphs

Abstract

Knowledge graphs (KGs) use triples to describe real-world facts. They have seen widespread use in intelligent analysis and applications. However, the automatic construction process of KGs unavoidably introduces possible noises and errors. Furthermore, KG-based tasks and applications assume that the knowledge in the KG is entirely correct, which leads to potential deviations. Error detection is critical in KGs, where errors are rare but significant. Various error detection methodologies, primarily path ranking (PR) and representation learning, have been proposed to address this issue. In this paper, we introduced the Enhanced Path Ranking Guided Embedding (EPRGE), which is an improved version of an existing model, the Path Ranking Guided Embedding (PRGE) that uses path-ranking confidence scores to guide TransE embeddings. To improve PRGE, we use a rotational-based embedding model (RotatE) instead of TransE, which uses a self-adversarial negative sampling technique to train the model efficiently and effectively. EPRGE, unlike PRGE, avoids generating meaningless false triples during training by employing the self-adversarial negative sampling method. We compare various methods on two benchmark datasets, demonstrating the potential of our approach and providing enhanced insights on graph embeddings when dealing with noisy KGs.